從真實影片案例學習 MiniMax H3 提示詞

DeCat · @DeCat2025

A timed five-shot prompt transforms a mountain into a guardian and a knight into its sword with explicit sound and continuity.

Snicklink · @snicklink

A locally generated interview short shares an eight-step FL2V setup, SLA attention and a separate CapCut finishing stage.

Prince does AI · @princedoesai

A creator presents a face-and-voice self-clone workflow with eight scenes, a two-minute result and a ComfyUI tutorial.

hvplan · @hvplan

An H3 creator assembles a short narrative around a 360 Orbit LoRA instead of presenting an isolated turntable test.

Cooking life hack · @HackCooking

A local H3 anime example includes dialogue, pairing stylized character action with a spoken scene.

さんぐらす❀ · @Ada_Lovelace49

An H3 360 Orbit LoRA experiment tests vertical rotation with two short Hatsune Miku clips.

みーみー｜AIアニメクリエイター · @mimi_N20

The same image and line compare H3 Max Turbo with two Kling models, with all voices replaced by the same ElevenLabs audio.

Kamimoto(かみもと) · @sep_is_heim

An H3 creator reports a fourteen-second clip in 98.8 seconds on RTX 4070, with faster reruns but weak speech quality.

Alexey Fateev · @superalesha

A melancholic MV credits H3 video and latent upscaling, local image models and YuE2 music on one RTX 3090.

あきばあい | 秋葉原IT戦略研究所 · @akibaAIilab

A Vidubus workflow documents an FP16 video VAE workaround for reported Mac MPS failures with an int8 VAE.

akiraxx · @theeakiraxx

An H3 ComfyUI test compares twenty and eight steps, Spectrum acceleration and an AudioRefine pass for speech distortion.

Ashraff Hathibelagal · @Hathibel

An H3 effects experiment describes lightning outlining a moving figure, glowing tattoos and timed floor ripples.

Modellix · @ModellixAI

A single-reference H3 dance example combines Y2K styling, choreography, close-ups and moving camera shots.

デジタルクリエイティブAI · @DCAI_TW

A local Ref2VA comparison reports faster four-step Turbo generation alongside weaker reference control and visible noise.

おくがわ🐟個人開発の魚 · @OnwUh12

An H3 comparison keeps video at six steps and tests six versus eight audio steps across three durations.

Amagami-lab | MOSOする人。 · @amagami_lab

A local H3 pipeline turns two character photos, a room photo and one brief into an eight-cut influencer short.

フースイ@AIイラスト · @ai_t2i_231

A five-second H3 animation documents a conversational Codex workflow that assembled English instructions behind the scenes.

きゅうす のAIとつれづれ · @Ai_Quse

A creator records DGX Spark memory and power during a long H3 workload, separating measurement from speed speculation.

SixteenGB · @SixteenGB

A 16 GB RTX 5070 Ti eGPU runs an H3 Ref2VA talking-cat example with int8, sparse attention and a four-step LoRA.

Glenn Williams · @GlennHasABeard

A public H3 prompt puts a tiny storm and house inside glass while a cloth wipe progressively clears the fog.

いるかいないか · @iruka_inainoka

Anima stills become five short H3 image-to-video shots, assembled with a Plenio piano track through ffmpeg.

DeCat · @DeCat2025

A public five-shot H3 prompt reveals a futuristic sky as removable panels inside an enormous mechanical sphere.

Feicai · @zhu185178

A six-stage short-drama tutorial compares H3 clip pricing with the time and iteration cost of finishing an entire film.

有趣的80后程序员 · @sunbaolong_2001

An eight-photo RefMod tutorial tests scene changes, wide shots, turnarounds and two-actor interactions with eight-step sampling.

_GAi · @_GAi92

A restaurant comedy uses a Mandarin fish-order joke in an H3 video shared with strong audience engagement.

Pierrick Chevallier | IA · @CharaspowerAI

An H3 motion-design example combines a moving character with oversized kinetic typography in the same composition.

烁皓 · @eternityspring

A period comedy joins 30-, 40- and 40-second Continuum runs using ten-second chunks and 22-frame handoffs.

码农暖爸 · @Delroy715

A same-seed H3 skin-rendering comparison combines two realism LoRAs, SigmaShift and a reported native 768p canvas.

有趣的80后程序员 · @sunbaolong_2001

A FastH3 Ref2V tutorial assembles two actors, clothing and props with eight-step sampling and two-stage latent upscaling.

crapthings · @crapthings

A 130-second zombie trailer compares H3 R2V footage above Three.js previs and documents reference and camera drift.